(question) MatMin instead of MatMul?

[TheNewSound]

Hi,

I'm relatively new to CUDA / Parallel programming and wondering if cutlass is the right library for me.

I want to create an algorithm which does a simple operation on a matrix A and its transpose B.

It's exactly the same as matrix-matrix multiplication, but not using the multiplying-operation on elements, but the min() operation on elements.

Is cutlass the right library for this? Can someone point me into the right direction?

[galv]

I'm not a CUTLASS developer, but, yes probably CUTLASS will work for your needs.

You'll need to implement a struct that uses min instead of multiplication for this template parameter to GemmConfig:

cutlass/cutlass/gemm/gemm_config.h

Line 49 in b5cab17

typename MultiplyAdd_,

[TheNewSound]

Thanks!! I actually just managed to rewrite my formula's to use Matrix Multiplication 🤓, which is more widely used (cuBLAS, upcoming NVIDIA's Tensor Cores etc.)

[TheNewSound]

With the new Tensor Cores upcoming, do I understand correctly that these Tensor cores have MultiplyAdd operations hardware-coded in them? Or is it also possible to use my Min operation on these new Tensor Cores (via cutlass?)

[galv]

MultiplyAdd operations are hardcoded into the tensor core hardware

…

On Thu, Aug 8, 2019 at 7:38 PM TheNewSound ***@***.***> wrote: With the new Tensor Cores upcoming, do I understand correctly that these Tensor cores have MultiplyAdd operations hardware-coded in them? Or is it also possible to use my Min operation on these new Tensor Cores (via cutlass?) — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#55?email_source=notifications&email_token=ABEL6UGOGPTXQKNQAJS43KTQDTKD7A5CNFSM4IJZEMY2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD35NOFI#issuecomment-519755541>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABEL6UAFMUDXVI35IYNHNRDQDTKD7ANCNFSM4IJZEMYQ> .

-- Daniel Galvez http://danielgalvez.me https://github.com/galv

[d-k-b]

Closing as not an issue. Please use the Discussions pages for further community input.